⚠️ Runtime not ready yet. These bundles use a new architecture (glm5_next: KDA linear attention + MLA/DSA hybrid + mHC) — vMLX Swift/Python runtime support is in active development and NOT released. Access is gated until it lands. Nothing loads these correctly today except the internal evaluation runtime they were built and measured with.

JANGQ-AI/GLM-5.3-Flash-JANG-MTP

GLM-5.3-Flash for 128 GB Macs — with the native multi-token-prediction layer preserved for self-speculative decode (~1.5–2× expected once runtime support lands).

A JANG bundle of zai-org/GLM-5.3-Flash — 300B-class MoE (288 experts, top-8 + shared) with KDA linear attention, sparse attention, and vision+video towers — quantized for Apple Silicon / MLX with a fully measured, per-unit dynamic bit allocation. Attention, routing, and all gating parameters are kept at 8-bit or full precision; every low-bit byte lives in the routed experts, placed by measurement.

Sibling bundle: GLM-5.3-Flash-JANG

Quality — measured, not estimated

15,850 teacher-forced positions on held-out prompts, versus the official FP8 release's logits (top-128, renormalized):

Bundle	Size	median KL	mean KL	p90 / p95 / p99	top-1	top-5	top-10
JANG-MTP	95.47 GiB	0.0977	0.552	1.57 / 2.63 / 5.74	78.1%	94.4%	96.7%
JANG (AR)	95.48 GiB	0.0885	0.529	1.50 / 2.56 / 5.64	78.7%	94.6%	96.9%

For calibration: our dots3-note release (280B at 94.6 GiB) ships at 79.2% top-1 — this model carries ~305B of routed experts with no fp16 embedding table to absorb signal, so these numbers are the honest physics of ~2.2 effective bits at this size.

How it compares (same protocol, same reference, same positions)

Quant (95 GiB class)	Size	median KL ↓	mean KL ↓	top-1 ↑	top-5 ↑	top-10 ↑
GLM-5.3-Flash-JANG (AR)	95.35 GiB	0.0885	0.52	78.7%	94.6%	96.8%
GLM-5.3-Flash-JANG-MTP	95.47 GiB	0.0977	0.55	78.1%	94.4%	96.7%
orcarouter GLM-5.3-Flash-MLX 2bit-lite	95.4 GiB	0.2122	0.83	71.4%	90.8%	94.3%

All rows: 15,850 teacher-forced positions vs the official FP8 release (top-128 renormalized KL). The orcarouter bundle was evaluated by loading its quantized weights natively — its exact shipped fidelity, no requantization. antirez's GLM-5.3-Flash-Q2.gguf (89.9 GiB) could not be measured: no public llama.cpp build (mainline or the open support PR) currently loads those files.

Calibration data

600,064 calibration tokens — 50% web text, 25% code, 15% multi-turn chat (incl. tool-call transcripts), 10% math/reasoning — with evaluation prompts drawn from a disjoint held-out tail. Activation-aware scaling and a per-channel importance refit are applied throughout; per-expert statistics cover 285–288 of 288 experts per layer.

What's in the bundle

• Vision + video: full tower (8-bit) + the consolidated image/video processor config.
• MTP: the native multi-token-prediction layer is preserved (shares the sparse-attention indexer). Draft depth requires a measured sweep on the target runtime.
• Thinking + agentic: thinking ON by default (the template force-opens <think>), reasoning efforts low / high / max (default max), clear_thinking=false preserves history thinking. Tool calls use GLM's XML dialect (<tool_call>name<arg_key>…<arg_value>…) — declared as tool_parser: glm_xml_args in the config; Hermes-style JSON parsers will not work.
• Self-describing quantization: per-module quantization block (bits + group size for all 626 quantized modules) in config.json.
• 128 GB Mac fit: ~95.5 GiB weights with a uniquely small cache footprint (fixed-size linear-attention state + compressed-latent KV ≈ 6 KB/token) — long contexts do not balloon memory.

Serving contract

• Sampling: temperature=1.0, top_p=0.95 (vendor defaults)
• EOS: [154820, 154827, 154829] · context: 1M native
• Reasoning: efforts low/high/max via reasoning_effort chat-template kwarg, default max; clear_thinking strips history thinking when true
• Tools: glm_xml_args dialect; tool results render as <|observation|><tool_response>…

Quantized and validated by Jinho Jang — eric@jangq.ai